Cambricon’s expansion from inference into training demonstrates how China and U.S. hyperscalers are reducing their dependence on Nvidia